lkcl (luke.leighton.delete@this.gmail.com) on July 28, 2022 3:38 pm wrote:
> Adrian (a.delete@this.acm.org) on July 25, 2021 9:16 pm wrote:
>
> > This "streaming" mode is actually exactly like the traditional vector computers have operated. For
> > example a Cray-1 had an apparent vector register width of 1024 bits, but the vector operations were
> > computed by a 64-bit pipelined ALU, in multiple clock cycles, i.e. "looping", like you say.
>
> the really nice thing about Cray-style Vectors is the "setvl" instruction.
> you request an arbitrary length (not limited to a power-of-two) and the
> hardware sets VL equal to as much as it can manage.
>
> what i *don't* know - and can't tell even by looking at the pseudocode linked from here:
> https://bugs.libre-soc.org/show_bug.cgi?id=893#c16
>
> is whether the Scalable Matrix Extension can set an arbitrary non-power-of-2
> dimension size. tracking through the pseudocode for CurrentVL to its definition
> goes through a cascading sequence of hoops of other definitions.
>
> can anyone tell for certain if SME can do matrix dimensions of say 7, or 21,
> or 35, as opposed to being specifically restricted to 2/4/8/16/32/... ?
>
> tia,
>
> l.
>

As far as I can see they have two predicates controlling which elements are active in a tile so yes they should be able to do that sort of thing.